Seeing History Unseen
Evaluating Vision-Language Models for WCAG-Compliant Alt-Text in Digital Heritage Collections
The paper Seeing History Unseen: Evaluating Vision-Language Models for WCAG-Compliant Alt-Text in Digital Heritage Collections is part of the proceedings of the Computational Humanities Research 2025 conference at the Luxembourg Centre for Contemporary and Digital History (C2DH) at the University of Luxembourg, published in the Anthology of Computers and the Humanities, Vol. 3, pp. 1147–1167. Slides for the accompanying short paper presentation are also available. The study evaluates images from the Open Research Data Platform of Stadt.Geschichte.Basel; all code, data, and survey materials are available in the companion repository.
Abstract
Digitized heritage collections remain partially inaccessible because images often lack descriptive alternative text (alt-text). We evaluate whether contemporary Vision-Language Models (VLMs) can assist in producing WCAG-compliant alt-text for heterogeneous historical materials. Using a 100-item dataset curated from the Stadt.Geschichte.Basel Open Research Data Platform—covering photographs, maps, drawings, objects, diagrams, and print ephemera across multiple eras—we generate candidate descriptions with four VLMs (Google Gemini 2.5 Flash Lite, Meta Llama 4 Maverick, OpenAI GPT-4o mini, Qwen 3 VL 8B Instruct). Our pipeline fixes WCAG and output constraints in the system prompt and injects concise, collection-specific metadata at the user turn to mitigate “lost-in-the-middle” effects. Feasibility benchmarks on a 20-item subset show 100% coverage, latencies of ~2–4 s per item, and sub-cent costs per description. A rater study with 21 humanities scholars ranks per-image model outputs; Friedman and Wilcoxon tests reveal no statistically significant performance differences, while qualitative audits identify recurring errors: factual misrecognition, selective omission, and uncritical reproduction of harmful historical terminology. We argue that VLMs are operationally viable but epistemically fragile in heritage contexts. Effective adoption requires editorial policies, sensitivity filtering, and targeted human-in-the-loop review, especially for sensitive content and complex figures. The study contributes a transparent, reproducible workflow, a small but representative evaluation set, and an initial cost–quality baseline to inform GLAM institutions considering AI-assisted accessibility at scale.